A Physicist Ran 36 AI Science Projects in 3 Months. The Hard Part Was Knowing Which Ones Mattered.

In a guest post published on Anthropic's research blog, Harvard physicist Matthew Schwartz describes what happened when he stopped trying to make Claude behave like a human collaborator and instead went hunting for problems shaped like the model's actual strengths. The output he reports is unusual in scale: 36 manuscripts across 18 fields with 19 coauthors in three months, drawn from roughly 400 candidate problems. The more interesting finding for anyone building agent systems is not the volume — it's where the pipeline kept breaking.

The reframe: stop fixing the model, change the problem

Schwartz's framing is borrowed from electrical engineering. He calls the gap between what scientists want and what current models do well an "impedance mismatch" — two systems that each function fine on their own but are badly matched, so most of what goes in never gets through. His earlier experiment, which he says involved Claude Opus 4.5 performing like a strong graduate student at twenty times the speed, still required correcting every sentence and pulling the model back from dead ends.

The pivot was to look for what he calls "Claude-shaped" problems: ones where the model's breadth of knowledge, coding ability and capacity to parse papers and data at speed do the heavy lifting, and where deep conceptual judgment is not the bottleneck. His starting point was the semi-numerical bootstrap in high-energy physics — a method that pins down scattering amplitudes by imposing physical constraints and then computing a handful of points to extreme precision rather than grinding out the integrals directly.

He argues this class of problem is unusually well suited to agents for three reasons: it draws on mathematics, physics and computer science that no single person has mastered; it requires heavy coding and algorithm development; and crucially, it is checkable. Anyone can verify the answer against the integral to as many digits as they like by running two scripts. That verifiability is the quiet load-bearing element of the whole story.

BootLoops is a harness, not a model

The tooling that emerged is called BootLoops, which Schwartz describes as a harness for the LLM in the same sense that Claude Code or Codex are harnesses. He says it is open source and usable with whatever model you prefer.

The early physics results: after porting methods scattered across Wolfram Language, C++, Python, Julia and papers with no code at all into a common framework, he reports 30 integrals run end to end — 15 reproductions of known results by the new method and 15 that he says had never been computed before. He notes the model reproduced results from one of his own papers in roughly 20 minutes, against code that took him weeks, and then told him his approach was inefficient.

From there the toolkit spread by mathematical analogy. The same integral structures turn up in Bayesian evidence calculations in population genetics; finite-field methods used for Feynman integral reduction map onto problems in evolutionary biology. That portability is the actual asset here, and it is a reminder that agents are often most valuable as translators between literatures rather than as originators inside one.

The expert was the bottleneck, every time

The most useful part of the post is Schwartz's account of being wrong about his own excitement. Inside his field he can judge whether a Claude result is good, and he says it often isn't. Outside it, he found himself agreeing with the model's self-assessment — and that worried him enough to recruit domain experts.

The pattern repeated. On neutral biodiversity theory in ecology, Claude solved an equation formulated by Rampal Etienne in 2005 that Schwartz says nobody had solved at scale for 20 years, and the analysis found that tree species composition at Barro Colorado Island in Panama shifts 4.5 times faster than neutral theory permits. Ecologist James O'Dwyer's response, as Schwartz reports it, was that the result would "be met with a shrug by many ecologists" — the qualitative point was already known. O'Dwyer's counterproposal, subtracting the neutral prediction and studying the residual, became the actual paper.

Same arc in population genetics. A 30-year-old integral solved, applied to gnomAD, and geneticist Michael Desai was impressed technically but unmoved scientifically. His redirection toward correlations between pairs of mutations on a chromosome led to an analysis of 5.7 billion nearby mutation pairs from the 1000 Genomes Project and evidence for gene conversion — a mechanism Schwartz says most analyses of linked genetic variation ignore.

The lesson generalizes well beyond science: an agent can produce correct, novel, technically impressive output that nobody in the target domain cares about. Taste did not come from the model. Schwartz says Claude gravitates toward old, highly cited, long-forgotten debates, and that asking it to focus on what is newly possible helps but does not make its judgment trustworthy.

The orchestration layer looks like ordinary infrastructure

The setup Schwartz describes is notably unglamorous: Claude Code sessions in terminals on Google Cloud VMs, linked to GitHub and Overleaf repos, one session per project plus a master session that coordinates, allocates compute and validates results. Background subagents run computations, with intermediate state written to markdown files.

Two details stand out. First, he uses subagents partly as blast radius control — he says he regularly hits the model's classifiers, and isolating that to one agent prevents a whole session from being corrupted. Second, context compaction kept destroying continuity on long projects, so he had the model periodically reorganize and consolidate its files to preserve the current plan. Both are workarounds for infrastructure limits, not intelligence limits.

His listed failure modes will be familiar to anyone shipping agents: the model declares victory prematurely (a proof that was complete "up to one unproven lemma" where the lemma was the entire proof); automated checks can't be trusted, so he asks to see plots; vague claims like "good agreement" need interrogation; and the default behavior is to grind through a multi-day calculation rather than build a tool that makes it take minutes. He also reports that time estimates were never reliable in either direction, and that the model has a taste for drama, once describing three days of work as a two-year campaign.

What the post does not establish

This is a first-person account published on the blog of the company whose model is the subject, and it should be read that way. Nothing here has cleared peer review; the economics work is described as an NBER working paper, and Schwartz says several of the highlighted results are still undergoing further exploration and verification. The post does not disclose compute costs beyond saying the projects were compute- and token-intensive, nor does it break down how the roughly 400 candidate problems narrowed to 36 manuscripts — a success rate under 10 percent, if those numbers are directly comparable, which the post doesn't confirm.

It is also worth noting what made this work: checkable outputs, pre-existing public datasets, and a principal investigator with enough expertise to recognize garbage in at least one domain and enough standing to recruit experts in others. Those conditions are not universal. Schwartz himself argues that AI sits inside the scientific loop rather than collapsing it, because most progress depends on real-world data that must be acquired, understood and checked.

What to watch

• Whether the BootLoops harness sees adoption outside Schwartz's own collaborations, and whether the claimed model-agnosticism holds up when someone swaps in a different frontier model.
• How the 36 manuscripts fare in peer review, particularly the ecology, population genetics and phylogenetics results where domain experts reshaped the questions after the fact.
• Whether the "checkable output" requirement proves to be the real gating factor — and what agentic science looks like in fields where no two-script verification exists.
• Cost transparency. Without disclosed compute spend, there is no way to judge whether this approach is reproducible by a typical academic lab or only by someone with vendor-scale access.
• Whether harness vendors absorb the workarounds Schwartz built by hand — context consolidation, adversarial referee sessions, per-agent failure isolation — into default product behavior.